A method is provided for predicting occurrence of a temperature rise event caused by a fire within an environment. The method comprises receiving temperature data captured by at least one temperature sensor 6, 8 for sensing an ambient temperature in the environment, processing the temperature data captured by the at least one temperature sensor in a previous window of time using a trained machine learning model 16 to determine a risk indication indicating risk of the temperature rise event occurring in a future window of time, and outputting a warning indication in dependence on the risk indication determined using the machine learning model. This is useful for providing firefighters with advance warning of dangerous temperature rises such as flashover events.
|
15. firefighter protection apparatus comprising:
at least one temperature sensor configured to sense an ambient air temperature within an environment surrounding the apparatus, where at least one of the at least one temperature sensor is provided on a portable or wearable device which portable or wearable by a firefighter; and
processing circuitry configured to:
process the ambient air temperature data captured by the at least one temperature sensor in a previous window of time using a trained machine learning model to determine a risk indication indicative of a risk of a temperature rise event, which comprises a rise in the ambient air temperature within the environment, occurring in a future window of time, wherein the risk indication is based on one or more predicted ambient air temperature values for the future window of time predicted using the trained machine learning model; and
when the risk indication satisfies a predetermined condition, control output of a temperature rise warning indication in dependence on the risk indication determined using the machine learning model; and
wherein the temperature rise warning indication comprises at least one of: an audible indication, a visual indication, and a vibrating indication; and
the processing circuitry is configured to process the ambient air temperature data captured by the at least one temperature sensor in successive sliding windows of time using the trained machine learning model, to determine the one or more predicted ambient air temperature values for corresponding future sliding windows of time.
1. A firefighter protection method for predicting occurrence of a temperature rise event caused by a fire within an environment, comprising:
receiving ambient air temperature data captured by at least one temperature sensor configured to sense an ambient air temperature within the environment, where at least one of the at least one temperature sensor is provided on a portable or wearable device which is portable or wearable by a firefighter;
processing the ambient air temperature data captured by the at least one temperature sensor in a previous window of time using a trained machine learning model to determine a risk indication indicative of a risk of the temperature rise event, which comprises a rise in the ambient air temperature within said environment, occurring in a future window of time, wherein the risk indication is based on one or more predicted ambient air temperature values for said future window of time predicted using the trained machine learning model;
when the risk indication satisfies a predetermined condition,
outputting a temperature rise warning indication in dependence on the risk indication determined using the machine learning model;
wherein the temperature rise warning indication comprises at least one of: an audible indication, a visual indication, and a vibrating indication; and
the processing comprises processing the ambient air temperature data captured by the at least one temperature sensor in successive sliding windows of time using the trained machine learning model, to determine the one or more predicted ambient air temperature values for corresponding future sliding windows of time.
2. The method of
3. The method of
at least one of said one or more predicted air temperature values exceeding a temperature threshold; and
a rate of change of predicted air temperature values with time exceeding a rate of change threshold.
4. The method
5. The method
6. The method
7. The method of
8. The method of
9. The method of
10. The method of
11. The method of
12. The method of
detecting one or more periods when an orientation of the body-worn device is outside at least one predetermined limit; and
excluding from the processing for determining the one or more predicted ambient air temperatures, temperature data captured by the remote temperature sensor during said one or more periods.
13. An apparatus comprising processing circuitry configured to perform the method of any of
14. A non-transitory recording medium storing a program comprising instructions to control a data processing apparatus to perform the method of
16. The method according to
|
This application is the U.S. national phase of International Application No. PCT/GB2017/051586 filed Jun. 2, 2017 which designated the U.S. and claims priority to GB Patent Application No. filed 1610022.4, filed Jun. 8, 2016, the entire contents of each of which are hereby incorporated by reference.
The present technique relates to the field of firefighting. More particularly, it relates to a method, apparatus and program for predicting an occurrence of a temperature rise event caused by a fire.
Safety is an important consideration for firefighters when attending to a fire. While current protective clothing for firefighters can deal with temperatures lower than around 300° C., temperatures during a fire can reach higher than this and can exceed 600° C. Above 300° C. charring of protective clothing begins and in the range 500 to 700° C. there is no chance of survival for any human beings in the vicinity of the fire. During the fire, the rate at which temperature increases can be highly non-linear and unpredictable, especially for a fire within an enclosed space such as a domestic room or office space. Due to the protective clothing worn by the firefighter it is difficult for the firefighter to sense the temperature increasing. The temperature can increase rapidly during the fire, especially in a condition known as “flashover” when combustible materials in the vicinity of the fire spontaneously ignite. It has been proposed to provide firefighters with a sensor for sensing ambient temperature which can trigger a warning if the temperature rises above a threshold. However, in practice by the time the temperature has risen above the threshold, the firefighter has little time to escape unharmed. Alternatively, the threshold could be set lower, but in this case the warning may be triggered some time before flashover actually occurs, and this may risk the firefighter leaving the scene of the fire prematurely reducing the chance of rescuing others from the fire. Therefore, a threshold-based approach may not provide an adequate warning of temperature rises.
At least some examples provided a method for predicting occurrence of a temperature rise event caused by a fire within an environment, comprising:
receiving temperature data captured by at least one temperature sensor configured to sense an ambient temperature within the environment;
processing the temperature data captured by the at least one temperature sensor in a previous window of time using a trained machine learning model to determine a risk indication indicative of a risk of the temperature rise event occurring in the environment in a future window of time: and
outputting a warning indication in dependence on the risk indication determined using the machine learning model.
At least some examples provide an apparatus comprising processing circuitry configured to perform the method discussed above.
At least some examples provide a program comprising instructions to control a data processing apparatus to perform the method described above.
At least some examples provide a firefighter protection apparatus comprising:
at least one temperature sensor configured to sense an ambient temperature within an environment surrounding the apparatus; and
processing circuitry configured to process the temperature data captured by the at least one temperature sensor in a previous window of time using a trained machine learning model to determine a risk indication indicative of a risk of a temperature rise event occurring in the environment in a future window of time, and to control output of a warning indication in dependence on the risk indication determined using the machine learning model.
Further aspects, features and advantages of the present technique will be apparent from the following description of examples, which is to be read in conjunction with the accompanying drawings, in which:
A method is provided for predicting the occurrence of a temperature rise event caused by a fire within an environment. The method comprises receiving temperature data captured by at least one temperature sensor configured to sense an ambient temperature within the environment, processing the temperature data captured by the sensor in a previous window of time using a trained machine learning model to determine a risk indication indicative of a risk of a temperature rise event occurring in the environment in a future window of time, and outputting a warning indication in dependence on the risk indication determined using the machine learning model.
Hence, a machine learning model can be trained (e.g. based on temperature data measured in a previous fire) to recognise patterns of previous temperature measurements which indicate that there is a risk of the temperature rise event occurring in a future window of time. Based on the risk identified by the machine learning model, a warning can then be output, for example to signal that the firefighter should leave the environment of the fire to preserve their own safety. By using a trained machine learning model to predict future risk of temperature rises based on temperature sensor data from a previous window of time, greater advance warning of temperature rise events can be provided.
The predicted temperature rise event could be any desired pattern of temperature rise to be detected, such as whether temperature has reached a given temperature threshold, or whether the rate of change of temperature rise is greater than a given amount, or some other condition of interest. However, the present technique is particularly useful where the predicted temperature rise event comprises a flashover event. Flashover is a condition during a fire when combustible materials within the environment spontaneously ignite because the air temperature in the room exceeds the auto ignition temperature of those materials. Flashover can pose a significant risk to a firefighter's safety since it can cause a sudden increase in temperature beyond the maximum temperature which can be withstood by the safety clothing worn by the firefighter, and so providing advance warning of a flashover event occurring can be valuable for ensuring the safety of the firefighters.
The processing using the trained machine learning model based on the temperature data captured by the sensors may be repeated over time based on different windows of sensor data, to provide continuing updates on the risk of temperature rise event occurring. Hence, the trained machine learning model may be supplied with temperature data captured in successive sliding windows of time, and process each window of temperature data to determine a risk indication for corresponding future sliding windows of time. For example, the temperature values may be captured at a certain refresh rate, and the most recent J values may be input to the machine learning model, which may predict the risk indication for a future sliding window of time corresponding to a certain number K of subsequent temperature measurements. The processing of a single window of sensor data by the machine learning model may be completed within one refresh period of the sensor data, so that each time new sensor data is received the processing is ready to process it. In this way the risk indication is repeatedly updated, and the warning indication is output if it is determined that there is sufficient risk of the temperature rise event occurring in the future window of time.
In some cases the risk indication may comprise an indication of a one of a plurality or types of risk class, e.g. whether the risk of the temperature rise event is considered to be low, medium or high. Hence, in some examples the risk indication could simply be a qualitative estimation of risk. In this case, the warning indication could be output when the risk indication indicates one of the predetermined subsets of types of risk class, e.g. if the risk is considered high. This can still provide sufficient warning to the firefighter.
Alternatively, the risk indication could be a more quantitative indication such as the probability of the temperature rise event occurring.
However, it can be particularly useful for the risk indication to comprise one or more predicted temperature values predicted for the future window of time. Hence, rather than a mere indication of risk, the trained machine learning model can predict the actual temperature values expected to occur in future based on the previously monitored temperature data for the previous window of time. Predicting actual temperature values can be useful because it allows additional information to be determined other than the risk indication, e.g. the expected time until the temperature rise event occurs could be predicted.
When future temperature values are predicted by the model, the method may include determining whether the predicted temperature values satisfy a predetermined condition, and outputting the warning indication if the predetermined condition is satisfied. For example the predetermined condition could be whether the predicted temperature values include at least one value exceeding a temperature threshold (e.g. 500° C.), or whether a rate of change of the predicted temperature values with time exceeds a rate of change threshold.
Various types of machine learning model could be used for predicting the risk of the temperature rise event. For example the trained machine learning model could be a supervised machine learning model such as an artificial neural network (ANN), hidden Markov model (HMM), or support vector machine, or an unsupervised machine learning model such as clustering models (KNN) or random forest algorithms. The model can be trained by providing training data comprising a series of temperature data recorded by sensors during a monitored fire.
For example, during a training phase (performed in advance, not at the scene of the fire when the method is used in the field), initial model parameters, such as weights or coefficients used for combining the input sensor data to calculate the risk indication, may be selected, and an iterative approach may be used for refining the model parameters based on the training data. In each iteration, the training data may be split into windows, and each window of training data may be processed using the model defined by the current model parameters to predict a risk indication (either a qualitative risk class, or future temperature values) for a corresponding subsequent window of time. The risk indication predicted for each window can then be compared with the actual risk evaluated from the subsequently captured temperature values in the corresponding subsequent window (or the future temperature values themselves), to calculate an error. An overall error score for the iteration can be derived from the errors calculated for each window of training data (e.g. by averaging the error for each window). The model parameters can then be adjusted and another iteration performed to see whether the error score increases or decreases. Subsequent adjustments to the model parameters used for later iterations can then be made that tend to cause the error score to decrease (e.g. using a gradient descent algorithm or other optimization algorithm), and the process is iterated until the error score is less than a given threshold (e.g. a prediction error of less than 0.7% could be required). At this point, the model parameters used for the final iteration are saved as the model parameters to be used when the technique is subsequently used in the field. Hence, the patterns to be identified for evaluating the risk indication can be learnt automatically by a supervised or unsupervised machine learning approach, rather than being defined in advance by a human designer.
While many types of machine learning model are available, it can be particularly useful to use an artificial neural network as the trained machine learning model because an artificial neural network is better able to generalise predictions even in the presence of noise in the input data, for example ignoring outliers, and is able to make graduated predictions which can better estimate the risk.
In some cases the same trained machine learning model may be used for fires in all types of environment. However, different types of environments may have different characteristics, e.g. different patterns or rates with which temperature rises during the fire. For example, a fire in a domestic setting may behave differently to a fire on board a ship or submarine within a metal compartment due to the different materials or dimensions used for those environments. To provide more accurate predictions for the type of environment being dealt with, the trained machine learning model may be selected from multiple models associated with the different types of environment. The multiple models could be of different types (e.g. one HMM and one ANN), or could be of the same type (e.g. ANN) but with different model parameters learnt based on different training data captured in the respective types of environment. The selection between the candidate models could be based on stored configuration data set statically in advance for a given device (e.g. a given device may be designated for use in domestic or marine environments respectively, but could be switched by updating the configuration data if it is reallocated for a different environment). Alternatively, the selection could be based on a user input. For example, when attending the firefighter (or an operation controller managing the rescue operation) may press a button or provide input (which could be sent from a remote device separate from the device held by the firefighter) specifying what type of environment is being attended to. Another option would be to provide a dynamic selection of the model to use depending on automated detection of the type of environment, e.g. based on image recognition or machine vision for detecting based on a camera image properties which indicate the type of environment, or based on GPS or another location sensor, which could allow detection of whether the firefighter is on land or at sea for example.
Nevertheless, in many cases a single trained machine learning model may be sufficient, especially if the firefighter is generally expected to attend a certain type of fire most often. Hence, some embodiments may use a single trained model for all conditions.
The one or more temperature sensors used to gather the temperature data for inputting to the machine learning model may be implemented in different ways. In some cases at least one static temperature sensor could be used which is disposed at a static location within the environment of the fire. For example, this could be useful for certain environments such as commercial buildings (e.g. offices), public buildings (e.g. libraries or museums) or industrial premises (e.g. factories) could be fitted with one or more temperature sensors at various locations, which in the event of a fire can provide temperature data for use by the fire services to predict whether a temperature rise event is likely in the future.
However, often firefighters may attend a fire within an environment which has not been instrumented with such temperature sensors. For example, it may be relatively unlikely that a domestic home may have such temperature sensors fitted. Therefore, it can be useful to use at least one body-worn temperature sensor which is disposed on a body-worn device within the environment. Hence, the firefighter can wear the body-worn device to carry their own temperature sensors into the environment of the fire, which gather temperature data to be input into the trained machine learning model for evaluating the risk of the temperature rise event occurring. For example, the body-worn device could comprise clothing worn by a firefighter, a helmet or face guard worn by the firefighter, or any other device or item which can be physically attached to the firefighter so as to move with the firefighter as the firefighter attends to the fire. However, it can be particularly useful for the temperature sensor to be mounted on a breathing apparatus used by the firefighter. Typically, firefighters entering a burning building may carry an oxygen supply to allow them to breathe in an environment containing potentially toxic fumes and low oxygen levels. Mounting the temperature sensor on the breathing apparatus ensures that the temperature sensor is unlikely to be forgotten (as the firefighter would generally carry the breathing apparatus), and also the breathing apparatus typically provides an area where the sensor can be mounted which is unlikely to get in the way of the firefighter's movement.
It can be useful for at least one temperature sensor to be provided for sensing the ambient temperature at a higher level in the environment than the level at which the firefighter is expected to be located, for example at ceiling height or another height above the level of the firefighter. This is because when a fire burns within a room, during an initial phase of the fire as the fire starts, hot gases and fumes from a burning object (such as a piece of furniture or waste bin) accumulate at the top of the room. As the fire progresses, more gases are released, and so the boundary between the hot gases above and cooler air below (known as the “neutral plane”) gradually descends down towards the floor. Hence, at a given height, the temperature will tend to increase with time as the neutral plane passes it, and temperatures recorded at higher levels will tend to increase earlier than temperatures recorded at lower levels. Therefore, by providing a sensor for sensing the temperature at a relatively high level in the environment, such as near the ceiling of a room, the trained machine learning model may be able to provide more advanced warning of temperature rise conditions at levels lower down, such as the shoulder height for the firefighter.
There are a number of ways in which ceiling temperature measurements can be made. In the case of static sensors provided in advance within the environment of the fire, the sensors can simply be positioned at the required height. For example, a thermocouple or other type of “proximity” type temperature sensor for sensing temperature in the immediate vicinity of the sensor itself could be used. Alternatively, the temperature at ceiling height can be sensed using a temperature sensor which is physically situated at a lower height, such as on the body of the firefighter. This can be done using a remote temperature sensor which can sense the ambient temperature at a location remote from the location of the temperature sensor itself. For example, the remote temperature sensor may comprise an infrared temperature sensor which senses temperature based on infrared radiation which is incident on the sensor from the remote location. All objects radiate electromagnetic radiation (known as black body radiation) with an intensity and wavelength distribution dependent on the temperature of the object. The infrared temperature sensor may have optics for focussing infrared radiation from a given region onto the sensor, and the temperature at that region can be deduced from the intensity and/or wavelength of the infrared radiation detected. Hence, for the case of body-worn sensors it can be particularly useful for the at least one temperature sensor to include at least one infrared temperature sensor or other remote temperature sensor to allow temperatures above the firefighter to be measured. In the case of an infrared sensor, the orientation with which the sensor is mounted on the body-worn device may be such that when in use, the sensor points upwards.
In some embodiments, the machine learning model may determine the risk indication based on temperature data from a single temperature sensor in some embodiments. This approach may be particularly effective if the single sensor senses temperature at ceiling height, since the ceiling temperatures may provide greater lookahead time for predicting rises at the firefighter's level.
However, the accuracy and lookahead time of the prediction can be improved when multiple temperature sensors are provided for sensing ambient temperature at different heights within the environment. This allows the correlation between the temperature rises at different heights to be used by the model to identify patterns likely to indicate a risk of a temperature rise event. For example, the temperature rise at the firefighter's level tends to lag the temperature rise higher up, so considering multiple sensors at different heights can improve the robustness of the prediction and reduce sensitivity to noise in the sensing of a single sensor. Again, there a different ways in which temperature sensors could be arranged to sense temperature at different heights. In a static case the temperature sensors may simply be positioned at different heights in the instrumented environment. In the body-worn case, the firefighter could carry a device with multiple proximity temperature sensors spaced apart from one another at different vertical levels when the device is worn. For example, the firefighter could carry some kind of frame or support with a section that extends upwards to carry the temperature sensor for sensing temperature at a higher level. However, this may restrict the movement of the firefighter in some circumstances, and so it may be less intrusive to the firefighter to provide a device which has at least one proximity sensor (such as a thermocouple) for sensing temperatures at the level of the firefighter, and at least one remote sensor which senses temperature at a higher level but is situated at the level of the firefighter. This allows a more compact device to be provided.
In the case where the body-worn device comprises at least one remote temperature sensor, the location at which the temperature is sensed by the remote temperature sensor may depend on the orientation of the firefighter as they move through the building. For example, an infrared sensor may have a limited field of view, and if the firefighter has to tilt or crawl along the ground, the field of view of the sensor may no longer be pointing upwards. To address this, the method may include detecting periods when an orientation of the body-worn device is outside at least one predetermined limit, and excluding from the processing for determining the risk indication temperature data which is captured by the remote temperature sensor during those periods. For example, one or more inertial sensors such as an accelerometer, magnetometer or gyroscope may be used to sense the orientation of the device carrying the remote temperature sensor. The bound for the acceptable orientation could be for example that the remote temperature sensors should be pointing within a certain threshold (e.g. 20 degrees) of the vertical, and when the orientation of the body-worn device is such that the remote temperature sensor is pointing more horizontally than this threshold then such temperature data can be excluded.
In the periods when the orientation is outside the bounds, excluding the temperature data from the remote temperature sensor could result in no prediction being made for those periods. Alternatively, a prediction can still be made by substituting the temperature data of the remote temperature sensor with dummy data (e.g. the temperature that was previously measured in a valid period when the orientation was within the permitted bounds) or with interpolated data derived from the previously monitored data. Also, as discussed above some embodiments may already predict future temperature values using the machine learning model itself, so in this case the previously predicted future temperature values from the model could be substituted for the sensor data obtained at corresponding timings, during periods when the orientation is outside bounds. This allows predictions to continue even during periods when the orientation of the firefighter makes the remote temperature sensor data unreliable. In practice, the firefighter is likely to return to an upright position after a relatively short period, so that a short period of using predicted values instead of actual temperature values may not greatly reduce the accuracy of the model's prediction.
Temperature sensors need not be the only types of sensors used as inputs to the machine learning model. The model may also use sensor data captured by at least one further sensor which is arranged to sense a property of the ambient environment other than temperature. The trained machine learning model can use that further sensor data in combination with the temperature sensor data to determine the risk indication indicating the risk of the temperature rise event occurring. For example, the further sensor could be a gas sensor for sensing presence of one or more predetermined gases in the environment (such as carbon monoxide, sulphur dioxide, nitrogen oxides, or chlorine for example, which tend to be generated during a fire), a humidity sensor for sensing humidity in the environment, and/or a pressure sensor for sensing air pressure in the environment. Any combination of one or more of these types of further sensors could be provided. By providing additional information relating to the onset of the fire, more accurate predictions can be made by correlating changes in temperature against changes in other factors within the environment being monitored. Again, the further sensors may be static sensors or body-worn sensors being carried by the firefighter.
The warning indication for warning the firefighter when the temperature rise event is likely can be generated in different ways. For example, the warning could be an audible indication (e.g. an alarm sound or other noise), a visual indication such as a flashing light or display with a warning image or text, and/or a vibrating indication such as a buzzing action on a device worn by a firefighter which can be felt to make the firefighter aware of the risk. In general, any way of signalling to the firefighter that there is a risk of an imminent temperature rise can be used. It can be particularly useful to output the warning indication via a wearable device, which may not be the same device as the one carrying the temperature sensors or the processing circuitry for running the machine learning model. For example the wearable device could be a wrist-worn device such as a watch or bracelet, a body-worn device such as a vest, badge or piece of clothing, or a head-worn device such as a hat, helmet, facemask or heads up display. The wearable device may be in communication with the device carrying the sensing and processing circuitry via any known technique which could be wired or wireless. However, a wireless communication may be preferred to reduce the restriction on the firefighter's movements. For example, Bluetooth®, Wifi®, Zigbee® or another wireless communication protocol may be used.
In addition to warning the firefighter, the warning indication may also comprise a signal transmitted to a remote command centre which is remote from the environment of the fire itself. This can provide warning to the operation controller or other party monitoring the rescue operation to signal when their firefighter may be potentially in danger. Hence, the device carried by the firefighter could have longer range communication capability such as a cellular radio transmitter to communicate via cellular networks to the remote command centre.
An apparatus may have processing circuitry for performing the method discussed above. For example, the apparatus may have processing circuitry comprising either bespoke hardware (e.g. an field programmable gate array) or generic processing hardware executing software, for implementing the trained machine learning model. The apparatus need not include the temperature sensors itself. For example, in the case of a building which is fitted with static sensors, the apparatus providing the processing of the sensor data using the machine learning model may be a central processor or server, which receives sensor data captured by external sensors and returns a signal to the firefighter indicating the risk indication or a signal for triggering a warning to the firefighter.
A computer program may be provided comprising instructions for controlling a data processing apparatus to perform the method of receiving the temperature sensor data, using the trained machine learning model to evaluate the risk, and outputting a warning indication (which could simply be a signal to another device carried by the firefighter). The program may be stored on a recording medium, which could be transitory or non-transitory.
In other examples, a firefighter protection apparatus may be provided, intended to be carried by the firefighters at a scene of a fire. The firefighter protection apparatus may comprise at least one temperature sensor for sensing ambient temperature within the environment of the apparatus, and processing circuitry for running the trained machine learning model, determining the risk indication using the model and the sensor data from the at least one temperature sensor, and controlling output of a warning indication based on the risk indication. This type of device can be more generally applicable to any setting of the fire, since the firefighter can carry their own temperature sensing and temperature rise prediction functionality with them as they enter the location of the fire. Although it would be possible to provide an apparatus which gathers the temperature sensor data and outputs it to a separate server or other central processor to calculate the risk using the machine learning model, by providing such functionality on the firefighter protection apparatus itself, this reduces the risk to the firefighter since even if communications between the firefighter and the central location go down, the firefighter still has local processing circuitry for calculating the risk indication and providing a warning.
The firefighter protection apparatus could be any device carried by the firefighter, but it can be particularly useful to be a breathing apparatus, or a module adapted for attaching to a breathing apparatus. For example, a module may be provided for attaching to existing breathing apparatuses, with the enclosed module including the one or more temperature sensors and the processing circuitry (such as a microprocessor or other chip) for running the trained learning machine model. Alternatively, an integrated product comprising the breathing apparatus hardware and the at least one temperature sensor and processing circuitry for running the model may be provided.
The following characteristics can be noticed in a room fire based on the NIST data:
On the other hand, the living room tests shown in
Despite a difference in the time taken in both the cases for a ceiling temperature reading to rise above 500° C., it is possible to predict a flashover condition if the ceiling temperature's rate of change is monitored in conjunction with the shoulder temperature behaviour.
The apparatus 2 also has communication circuitry 24, 26 for communicating with other devices, including long range communication circuitry 24, such as a radio transmitter for communicating via cellular networks, for communicating with a remote server or other device used by an operation commander to control or oversee a number of firefighters dealing with the fire. The long range communication could be via a mesh network. Also, short range communication circuitry 26 may be provided for communicating locally with other devices held by the firefighter, such as a camera (not shown in
While
The wearable device 30 may for example be a watch, bracelet or wristband, piece of clothing, helmet or facemask or heads up display worn by the firefighter, and can be used to provide feedback to the firefighter. When the machine learning model running on the processing circuitry 4 predicts that there is a risk of the temperature rise event occurring, a signal is transmitted to the wearable device 30 by the short range communication circuitry 26, which triggers the wearable device 30 to output a warning. For example, the wearable device may include a display 32 for displaying text or images to the firefighter, an LED 34 other warning light for lighting up or flashing on and off to provide a warning, a speaker 36 for providing an audible warning, or a vibrating element 38 (e.g. a buzzer or piezoelectric element) for providing a tactile alarm to the firefighter which can be felt by the firefighter. The device 30 also includes communication circuitry 40 for communicating with the short range communication circuitry 26 in the firefighter protection apparatus 2, and processing circuitry 42 for processing signals received by the communication circuitry 40 and controlling the operation of the output devices 32 to 38 accordingly. It will be appreciated that a particular implementation of the wearable device 30 need not have all of these elements 32, 34, 36, 38. For example, a bracelet could be provided with an LED 34 and a vibrating element 38, but might not have the speaker, while a headset could be provided with a heads up display 32 and speaker 36, but no LED or vibrating element. Hence, the particular form of the alarm provided to the wearer may depend on the type of wearable device. Also, some implementations may use a combination of two or more types of wearable device (e.g. both a headset and a wrist watch).
One particularly useful example may be to provide a bracelet to be worn inside the firefighter's protective clothing, providing a warning using a vibrating element 38 which can be felt by the firefighter, as well as a smart watch or other wrist-worn device to be worn outside the protective clothing, comprising a display 32 for displaying the warning indication and optionally other information to the firefighter. Providing the inner bracelet increases the likelihood that the warning is perceived quickly by the firefighter, as it does not require the firefighter to be looking at the smart watch, but providing the additional watch allows further information to be provided, such as an indication of the temperature monitored, estimated time to flashover, degree of risk detected, etc.
While
In one particular embodiment, the processing circuitry 4 may comprise a chipset, running on an embedded ‘system on module’ (SOM), in essence, a compact stand-alone computer. The SOM is mounted on a printed circuit card, inside a rugged enclosure 2 which will also take various sensor inputs to collect and process data on the SOM. The core hardware on the breathing apparatus set also provides connectivity over wireless and cellular networks, allowing information to be transmitted from the breathing apparatus set module 2, including a GPS derived location for example. The shorter range wireless network communications are used to communicate with separate devices near to the breathing apparatus set module 2, e.g. a helmet mounted camera to provide video data, or a paired Bluetooth device e.g. a smart watch, to provide local feedback of information to the individual firefighter. The longer range cellular communications can be over both public and private cellular networks, to provide back-haul to a cloud-based server. This server could be accessed by an incident commander on a separate device, e.g. tablet computer, in order to view the information being generated from each of his/her firefighters. The software may use a commercial off-the-shelf operating system (e.g. Android®) running applications implementing the machine learning model, which may be written in a programming language such as Java. The SOM takes input data from compact, non-contact, infrared sensors, which also have a local thermocouple in the sensor head. The sensors may for example be commercial off-the-shelf CT-SF22-03 sensors provided by micro-epsilon. The non-contact infrared sensors comprise a thermopile coupled to a set of optics to provide a focus on a heated object. The sensors are mounted vertically to measure temperatures above the firefighter (i.e. the ceiling in most building environments), alongside the thermocouple measurement which is at shoulder height. The temperature data is fed into an artificial neural network, which has been trained from a comprehensive time series data set from the National Institute of Standards and Technology (NIST), based upon a house fire experiment. This experiment measured the development of temperatures and heat fluxes at different heights, inside different rooms, as the flashover condition develops. The artificial neural network described previously has been trained to sample data and predict the rate of temperature rise, and has been successful in predicting a flashover condition 27 seconds before occurrence. The local temperature and warning can be communicated to the firefighter via a paired device e.g. smart watch; temperature being given as a direct numerical value, while the warning indication can be colour-coded (e.g. green through to red) based upon a number of different classes, relating to the progression of the flashover condition. This data can also be shared by the longer range communications back to the cloud server so as to provide instant commander with this information.
It will be appreciated that
As shown in
Based on the predicted window of future temperature values, the processing circuitry 4 can determine whether any of these values meet a given condition for triggering a warning indication. For example, the processing circuitry 4 may determine whether any of the predicted temperature values exceed a given threshold (e.g. 500° C. for detecting a flashover event), or whether a rate of change of the predicted temperature values exceeds a given threshold. If the trigger condition is satisfied, a warning signal is sent to the wearable device 30, by the processing circuitry 4 controlling the short range communication circuitry 26 to signal a suitable control value to the wearable device 30.
As time progresses, the sliding past and future windows move on in time as more temperature data is received. For example, following the calculation shown in
The neural network comprises an input layer 100 comprising a number of nodes 102. Each of the temperature values Tth-4 to Tth0 and TIR-4 to TIR0 captured by the respective sensors 8, 6 during the sliding past window of time is input to a corresponding one of the nodes 102 of the input layer 100. If a further (non-temperature) sensor 10 is provided in the apparatus 2 then further nodes may be provided for receiving the sensor data from the further sensor 10. The input layer 100 may also have one or more additional nodes 103 which receive a constant value b which is independent of the captured sensor data (the constant value b may be one of the parameters which is learnt during the training phase).
The network also comprises at least one hidden layer 104 of nodes 106. In each hidden layer 104, each node 106 receives the values represented by each node 102 of the preceding layer. In this particular example, there is only one hidden layer 104, so each node 106 receives the values input at each node 102 of the input layer 100, but if multiple hidden layers 104 were provided then the first hidden layer would receive values from the nodes of the input layer 100, and subsequent hidden layers 104 would receive the values output by nodes 106 of a preceding hidden layer. Each node 106 of the hidden layer 104 outputs a value f(x0, x1, . . . , xN, w0, w1, . . . , wN) obtained as a function of all the inputs to that node and a corresponding set of weights w learnt during the training phase. Separate sets of weights may be defined for each node of the hidden layer 104 and for a given node 106 of the hidden layer, separate weights may be defined for each input x. The weights to be used may be defined by the stored model coefficients 16 in the data storage 14. The stored model coefficients 16 may also specify other parameters defining the particular function to be used.
For example, the output of a given node 106 of the hidden layer 104 could be determined according to the following equation:
f(x0 . . . xN,w0 . . . wN)=σ(t) (1)
where t is the weighted sum of the inputs to the node:
and σ(t) is a sigmoid function used as an activation function for controlling the relative extent to which each node 106 of the hidden layer 104 affects the final prediction:
β is a slope parameter defined in the model coefficients 16. The slope parameter β could be the same for all nodes or could be different and learnt for each node as part of the training phase. It will be appreciated that other activation functions could be used to control the extent to which the weighted sum of inputs affects the output, but the sigmoid function can be particularly useful because its derivative is easy to calculate which makes some training algorithms for updating the weights w easier to implement. It will be appreciated that equations 1 to 3 above are just one example of a possible function for determining the output of each node.
The final layer of the artificial network is an output layer 108 comprising a number of nodes 110 whose outputs represent the predicted temperature values T1 to T10. As for the hidden layer 104, each output node 110 receives the values calculated for each node 106 of the preceding layer 106, and calculates its output as a function of its inputs x and weights w, for example according to equations 1 to 3 shown above.
To train the network, temperature sensor data measured during a fire in a real environment (such as the NIST data mentioned above) is partitioned into windows of time. An initial set of weights w and other model parameters (such as the constant b or slope parameter β mentioned above) is selected, and each window of training data is input to the network and processed using the current model parameters to predict future temperatures is assumed way to shown in
Appendix A below shows results of applying the trained artificial neural network to the NIST living room data to test the model's predictive ability. For conciseness, Appendix A does not show all the past temperature values used for the prediction, and each line of Appendix A simply shows the “current temperature” and the 10 predicted temperatures predicted for the next 30 seconds. In this example, the neural network was trained based on the NIST data for the sensor at 7.42 foot in the living room data set, and tested on the corresponding height data from the dining room data set. The refresh period of the temperature sensors is 3 seconds and so each line of Appendix A represents a prediction made 3 seconds after the previous prediction. As shown in Appendix A, the current temperature is initially around 15-16° C., representing the room temperature before the fire is lit. Around lines 34-39 of Appendix A the temperature starts to increase due to the fire.
Line 56 of Appendix A is the first time that one of the predicted temperatures exceeds 500° C., signalling that a flashover event is predicted. At this point, the warning may be triggered to the firefighter. The actual temperature at shoulder height reaches 500° C. at line 70 of Appendix A. Hence, it took 14 time-steps between the model predicting flashover, and the real flashover event being encountered. As the refresh rate in this example is 3 seconds, this means that the model provided 42 seconds' warning of the imminent flashover. At the time that the warning was provided (line 56), the current temperature at shoulder height was still 269° C., which could still be safely withstood by the firefighter's protective clothing. Hence, this shows how the trained model can successively provide sufficient warning of an imminent flashover event.
FIG.
Thermocouple
Height from Floor (m)
Height from Floor (ft)
8
TC0
2.26
7.42
9
TC1
1.98
6.5
10
TC2
1.68
5.5
11
TC3
1.37
4.5
12
TC4
1.07
3.5
13
TC5
0.76
2.5
14
TC6
0.46
1.5
15
TC7
0.15
0.5
In each graph, line A represents the actual temperatures recorded by the corresponding thermocouple in the NIST dining room data set, and line B represents the temperatures predicted by the trained artificial neural network. When standing upright, a fire-fighter's standard height is presumed less than or equal to 6.5 ft, so sensors TC1-TC7 may all represent heights where temperature increases may pose a risk to the firefighter. However, fire-fighters do not always stand upright hence in a prone position the height is approximated at 0.76 m/2.5 ft, so the data from sensor TC5 may be of interest for this. For the particular dataset used, the actual flashover (based on the top-most TC0 temperature sensor) started at 3 minute-30 second time. However, similar temperature flashover condition for the sensor mounted at a standard fire-fighter height (1.98 m/6.5 ft) was reached in the next 3 seconds, and as the neutral-plane moved further below, for a prone-height (0.76 m/2.5 ft), similar flashover conditions were reported by the TC6 sensor in the next 1 minute-18-seconds. Based on
As shown in
As shown in
As shown in
As shown in
As shown in
As shown in
As shown in
As shown in
The results of these tests are summarised in the table below:
Time temperature
Time at which the
Original
Time taken for the
increased above
RoT algorithm
temperature
actual flashover to
Sensor
100-degrees
predicted the
at the
be reached after
height
before RoT Tf
flashover
prediction of
prediction
(m/ft)
prediction
temperature Tf
Tf (° C.)
(seconds)
2.26/7.42
1 minute
2 minutes
322
36
3 seconds
54 seconds
1.98/6.5
1 minute
2 minutes
265
36
3 seconds
57 seconds
1.68/5.5
1 minute
3 minutes
274
24
21 seconds
15 seconds
1.37/4.5
72 seconds
3 minutes
306
18
30 seconds
1.07/3.5
33 seconds
3 minutes
214
30
30 seconds
0.76/2.5
21 seconds
3 minutes
357
60
35 seconds
0.46/1.5
2 minutes
6 minutes
235
Flashover not
(false
54 seconds
39 seconds
reached
positive
Max Tf = 350
case)
reached in 33 seconds
0.15/0.5
1 minute 30
Flashover not
Flashover not
Flashover not
seconds
reached
predicted
reached
Hence, the actual flashover prediction ranged from 18 to 60 seconds with the longest predicted case recorded at the height of 0.76 m/2.5 ft. However, this height reported the fastest RoT from normal conditions of approximately 100-degree-celsius to prediction within 21 seconds. On 1.98 m/6.5 ft height, the flashover condition as predicted in 36 seconds with the temperature reaching the algorithmic prediction stage (3rd column) of 265 degrees Celsius in 1 minute 3 second duration. The duration shown in Column 2 can be considered as a “warning segment” which, on the average took approximately 1 minute 7 seconds for all 7 sensor predictions. For the lowest-height temperature sensor, the flashover was never predicted. However, in this case, the maximum temperature never went beyond 166 degrees Celsius. The second-last row showing the sensor mounted at 0.46 m/1.5 ft does however show a false positive as it predicted flashover closer to the 7th minute mark. However, the temperature only reached 350 degrees Celsius in the next 33 seconds before reducing again. Based on data from all the sensors, it can be understood that the majority of sensors predicted an impending flashover with a high accuracy. The earlier prediction however still came from the highest sensor. Nonetheless, the second and third highest sensors were still able to warn in ample time (in both cases the prediction was made before the 3 minutes 30 seconds time at which flashover occurs at ceiling height). This shows that while using ceiling height temperature measurements can provide greatest warning time, it is also possible to perform the prediction only using sensors mounted lower down on the body of the fire-fighter.
At step 202 the processing circuitry 4 determines, based on the orientation sensor 22, whether the orientation of the device is outside certain bounds. If so, at step 204 sensor data from the infrared sensor 60 is excluded from subsequent calculation or replaced with estimated sensor data such as sensor data obtained based on the interpolation of previously detected values when the orientation was within bounds, or previously predicted sensor data 20 determined using the model for the corresponding moments in time. On the other hand, if the orientation was within the permitted bounds then step 204 is omitted.
At step 206 the sensor data (with some of the data excluded or replaced as necessary in step 204) is input to the trained machine learning model which processes the data to evaluate each node of the network based on the coefficients or weights 16 learnt during training. The model outputs the predicted risk class or predicted temperature values for the future window of time. At step 208 the processing circuitry 4 determines whether the predicted risk class or temperature values meet certain warning criteria and if so then at step 210 a warning is output (either directly by the firefighter protection apparatus 2 itself, or via the short range communication 26 to an associated wearable device 30). A warning can also be sent via long range communication circuitry 24 to a remote device such as the operation commander's server or tablet. If the risk indication does not meet the warning criteria at step 208, then step 210 is omitted and no warning is output. At step 212 the temperature sensor data is updated with the next reading(s), to proceed to the next sliding window of time, and at step 202 the processing continues for the next window. Hence, the method is repeated again and again for successive windows of sensor data.
In the present application, the words “configured to . . . ” are used to mean that an element of an apparatus has a configuration able to carry out the defined operation. In this context, a “configuration” means an arrangement or manner of interconnection of hardware or software. For example, the apparatus may have dedicated hardware which provides the defined operation, or a processor or other processing device may be programmed to perform the function. “Configured to” does not imply that the apparatus element needs to be changed in any way in order to provide the defined operation.
Although illustrative embodiments of the invention have been described in detail herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various changes and modifications can be effected therein by one skilled in the art without departing from the scope and spirit of the invention as defined by the appended claims.
An example of a training process for updating the weights of an artificial neural network based on training data is shown below. It will be appreciated this is just one example of a training process, and other techniques could be used.
The process of back propagation undergoes a weight optimisation process to allow the underlying neural network to learn to map arbitrary inputs to outputs. The training process is an iterative process, with each iteration including a forward pass propagation portion where the network using a given set of weights calculates predicted temperatures based on the input training data and compares this with the actual temperatures for the corresponding timings to determine errors, and a backward error propagation pass where the detected errors between the predicted temperatures and the actual temperatures are used to update the weights.
Forward Pass Propagation
The steps for the forward pass calculation are given below:
The process shown in (1)-(3) is to be repeated again for input 2 as follows:
The above process is to be repeated for the output layer neurons as follows:
The total error for the first epoch would therefore be calculated based on (10) and (14) and the equation given below:
In (15), the original values are those shown in
The Backward Error Propagation Pass:
The goal of back propagation is to update each weight in the network so that actual output value is closer to the target value. This results in minimising error for each output neuron and subsequently for the entire ANN.
Consider weight wh101, now the objective is to find out how much change in this weight would change (reduce) the total error. This can be formulated as:
Equation (19) can be understood as the partial derivative of Etotal with respect to wh101 which is also called the “gradient with respect to wh101”. Applying chain rule on (19) would lead to equation
The partial derivative of the logistic function derivative is calculated as the output multiplied by 1 subtracting the output:
Based on (25), the total net input of o1 change with regards to wh101 can therefore be calculated as follows:
Putting all the values together based on equation (20):
Now, in order to decrease the error, the above value in (27) is subtracted from the current weight and optionally multiplied by a learning rate η which is currently set at 0.5 as follows:
The other weights can also be calculated in a similar fashion. Once all the weights have been updated, these weights are used in place of the older weights which is expected to reduce the error in the subsequent epochs. With the repetition of the abovementioned process for extended epochs (iterations), the overall error tends to drop to fractionally small values, at which point the network model is considered trained to be used against unseen data in order to predict expected outcomes.
Patent | Priority | Assignee | Title |
11751360, | Mar 17 2020 | International Business Machines Corporation | Intelligently deployed cooling fins |
Patent | Priority | Assignee | Title |
5973602, | Apr 30 1993 | COLE, BOYD F | Method and apparatus for monitoring temperature conditions in an environment |
6118382, | Oct 30 1997 | MSA Technology, LLC; Mine Safety Appliances Company, LLC | System and method for alerting safety personnel of unsafe air temperature conditions |
6417774, | Oct 30 1997 | Mine Safety Appliances Company | System and method for identifying unsafe temperature conditions |
6472988, | May 19 1998 | SCOTT TECHNOLOGIES INC | System for monitoring wearers of protective respiratory equipment |
6700497, | Oct 30 1997 | Mine Safety Appliances Company | System and method for identifying unsafe temperature conditions |
6868697, | Nov 19 2002 | Samuel Aaron Licensing, LLC | Engagement set with locking arrangement and rear crossover configuration |
6995665, | May 17 2002 | Mine Safety Appliances Company | System and method for identifying, monitoring and evaluating equipment, environmental and physiological conditions |
7073351, | Nov 19 2002 | Samuel Aaron Licensing, LLC | Engagement set with locking arrangement and rear crossover configuration |
7287400, | Nov 19 2002 | Samuel Aaron Licensing, LLC | Engagement set with locking arrangement and rear crossover configuration |
7676280, | Jan 29 2007 | Hewlett Packard Enterprise Development LP | Dynamic environmental management |
8085144, | Jul 02 2002 | Mine Safety Appliances Company | Equipment and method for identifying, monitoring and evaluating equipment, environmental and physiological conditions |
8400292, | Mar 01 2010 | CommScope Technologies LLC | System and method for location of mobile devices in confined environments |
8554703, | Aug 05 2011 | GOOGLE LLC | Anomaly detection |
8638214, | Mar 01 2010 | CommScope Technologies LLC | System and method for location of mobile devices in confined environments |
9000911, | Mar 01 2010 | CommScope Technologies LLC | System and method for location of mobile devices in confined environments |
9625324, | Apr 08 2016 | Thermal monitoring and estimating system and method | |
20020135488, | |||
20030058100, | |||
20030065409, | |||
20030158635, | |||
20040004547, | |||
20040131498, | |||
20050001728, | |||
20060006997, | |||
20060125623, | |||
20060270471, | |||
20070080817, | |||
20070103292, | |||
20070176770, | |||
20070205886, | |||
20070205903, | |||
20070241261, | |||
20080106437, | |||
20080146895, | |||
20080221741, | |||
20090184244, | |||
20090207852, | |||
20100081411, | |||
20100219956, | |||
20100315228, | |||
20110042109, | |||
20110140913, | |||
20110210843, | |||
20110265706, | |||
20120146787, | |||
20130147621, | |||
20130300535, | |||
20130335235, | |||
20130342361, | |||
20140203938, | |||
20140373601, | |||
20140375459, | |||
20150097687, | |||
20150112883, | |||
20150119079, | |||
20150124850, | |||
20160006953, | |||
20160049071, | |||
20160081415, | |||
20160089552, | |||
20160124475, | |||
20170169683, | |||
20170206464, | |||
20170251933, | |||
20170261380, | |||
20170278375, | |||
20170330437, | |||
20180233019, | |||
20180293864, | |||
20190012904, | |||
EP801368, | |||
EP1296745, | |||
GB2427732, | |||
JP2004102634, | |||
WO2010034547, |
Executed on | Assignor | Assignee | Conveyance | Frame | Reel | Doc |
Jun 02 2017 | STS Defence Limited | (assignment on the face of the patent) | / | |||
Dec 07 2017 | YUSUF, SYED | STS Defence Limited | ASSIGNMENT OF ASSIGNORS INTEREST SEE DOCUMENT FOR DETAILS | 044353 | /0218 | |
Dec 07 2017 | GARRITY, DAVID | STS Defence Limited | ASSIGNMENT OF ASSIGNORS INTEREST SEE DOCUMENT FOR DETAILS | 044353 | /0218 |
Date | Maintenance Fee Events |
Dec 11 2017 | BIG: Entity status set to Undiscounted (note the period is included in the code). |
Nov 15 2023 | M1551: Payment of Maintenance Fee, 4th Year, Large Entity. |
Date | Maintenance Schedule |
May 19 2023 | 4 years fee payment window open |
Nov 19 2023 | 6 months grace period start (w surcharge) |
May 19 2024 | patent expiry (for year 4) |
May 19 2026 | 2 years to revive unintentionally abandoned end. (for year 4) |
May 19 2027 | 8 years fee payment window open |
Nov 19 2027 | 6 months grace period start (w surcharge) |
May 19 2028 | patent expiry (for year 8) |
May 19 2030 | 2 years to revive unintentionally abandoned end. (for year 8) |
May 19 2031 | 12 years fee payment window open |
Nov 19 2031 | 6 months grace period start (w surcharge) |
May 19 2032 | patent expiry (for year 12) |
May 19 2034 | 2 years to revive unintentionally abandoned end. (for year 12) |